Transformers in surgical artificial intelligence: A domain-stratified, study-level narrative review

Abstract

Objective

The objective was to synthesize study-reported outcomes of transformer-based artificial intelligence systems in surgical domains and describe settings where they appear advantageous relative to nontransformer models and human benchmarks.

Methods

We searched major databases: PubMed, Embase, IEEE Xplore, ScienceDirect, Google Scholar, arXiv, and Cochrane Library. Eligible studies evaluated transformer architectures in surgical/perioperative contexts (medical imaging, workflow recognition, prognosis-related modeling, or education) and reported quantitative outcomes. Because of heterogeneous tasks/metrics, we performed a domain-organized narrative synthesis. Where the same study reported transformers and nontransformers on the same dataset/metrics, we computed within-study deltas (Δ = Transformers – Nontransformers) and summarized medians and interquartile ranges alongside vote counts (T>NT/tie/T<NT). No cross-study pooling or hypothesis testing was performed.

Results

Paired comparisons favored transformers in medical imaging for 15 of 20 (75.0%) with median Δ +1.13 percentage points (interquartile range, 3.70) and in workflow recognition for 28 of 34 (82.4%) with median Δ +1.75 percentage points (interquartile range, 3.28). Prognosis had sparse paired data (n = 1; Δ +3.0 percentage points; illustrative). Education favored transformers overall in 5 of 6 (83.3%) paired comparisons, driven by surgery time prediction; diagnostic education tasks were mixed. Reported advantages were task dependent and dataset specific; gains were typically single-digit percentage points in like-for-like settings.

Conclusions

Transformers frequently match or exceed nontransformer baselines in surgical imaging and workflow tasks, with promising, yet heterogeneously reported, signals in prognosis and education. Translation to dependable clinical/educational impact will require standardized benchmarks, external/prospective validation, transparent comparator reporting (including human baselines), and deployment studies that address real-time operating room constraints and fairness across patient and learner groups.